Unlock: The Loop Inside the Model

A language model can apply the same learned block several times to an evolving continuous state before it emits a word. That makes inference depth a dial, with no growth in unique parameters. What it does not yet show is that silent, continuous recurrence beats tokens, pauses, or search under matched budgets. This page states the mechanism, separates the three loops that get called one thing, and writes down the experiment that would settle the open claim.